8

Evaluation of Anomaly Detection

Capability for Ground-Based
Pre-Launch Shuttle Operations
Rodney A. Martin, Ph.D.
NASA Ames Research Center
Mail Stop 269-1
Moffett Field, CA 94035
U.S.A.
1. Introduction
This chapter will provide a thorough end-to-end description of the process for evaluation of
three different data-driven algorithms for anomaly detection to select the best candidate for
deployment as part of a suite of IVHM (Integrated Vehicle Health Management)
technologies. These algorithms will be evaluated based upon their capability for robustly
detecting incipient faults or failures in the ground-based phase of pre-launch space shuttle
operations, rather than based on system certifiability as performed in previous studies
(Schwabacher & Waterman, 2008). Robust detection will allow for the achievement of prespecified minimum false alarm and/or missed detection rates in the selection of alert
thresholds. All algorithms will be optimized with respect to an aggregation of these same
criteria. The final results will also include a formal cross-validation procedure, which will be
used to perform optimization and alert threshold selection.
The data-driven algorithms to be evaluated in this study were deemed to be sufficiently
mature for consideration as viable candidates for demonstration during the launch of Ares IX. This launch represents the first test flight of Ares I, which will be the successor to the
Space Shuttle for NASAs Constellation program. Our study relies upon the use of Shuttle
data to act as a proxy for and in preparation for application to Ares I-X data, which uses a
very similar hardware platform for the subsystems that are being targeted (TVC Thrust
Vector Control subsystem for the SRB (Solid Rocket Booster)).
Data-driven algorithms are just one of three different types being deployed, the details of
which were presented in previous work (Iverson et al., 2009); (Schwabacher & Waterman,
2008). The other two types of algorithms being deployed include a rule-based expert
system, and a model-based system. Within these two categories, the deployable
candidates have already been selected based upon non-quantitative factors such as flight
heritage and system certifiability. For the rule-based system, SHINE (Spacecraft Health
Inference Engine) has been selected for deployment, which is one of many components of
BEAM (Beacon-based Exception Analysis for Multimissions) as described in its debut article
(Mackey et al., 2001). Other components of BEAM include various data-driven algorithms.
BEAM is a patented technology developed at NASAs JPL (Jet Propulsion Laboratory) and
Source: Aerospace Technologies Advancements, Book edited by: Dr. Thawar T. Arif,
ISBN 978-953-7619-96-1, pp. 492, January 2010, INTECH, Croatia, downloaded from SCIYO.COM

www.intechopen.com

142

Aerospace Technologies Advancements

SHINE serves to aid in the management and identification of operational modes. For the
model-based system, a commercially available package developed by QSI (Qualtech
Systems, Inc.), TEAMS (Testability Engineering and Maintenance System) was highlighted
in work subsequent to its debut (Cavanaugh, 2001), and has been selected for deployment to
aid in diagnosis. In the context of this particular deployment, distinctions among the use of
the terms data-driven, rule-based, and model-based, can be found in the previously
cited paper by Schwabacher and Waterman.
Although there are three different categories of algorithms that have been selected for
deployment, our main focus in this chapter will be on the evaluation of three candidates for
data-driven anomaly detection. The main thrust of the chapter is to provide instructive
coverage on the topics of algorithmic optimization and alert threshold selection for the
candidate data-driven algorithms. These algorithms will be optimized and compared using
the AUC (Area under the ROC (Receiver Operating Characteristic) curve), which represents
overall classification discriminability. The resulting optimized algorithmic parameters can
then be used to perform threshold or alert selection based upon the corresponding ROC
curve. This allows for the demonstration of a robust anomaly detection capability when
performing alert threshold selection, which is based upon specified minimum false alarm
and/or missed detection rates. As a practical measure we will also present a performance
comparison among the candidate data-driven algorithms by evaluation of their
computational complexity as well as the AUC. The final results will also include a formal
cross-validation procedure, which will be used to provide uncertainty bounds for the
optimization and alert threshold selection processes.

2. Motivation & background

Some insightful previous studies (Fragola, 1996); (Pat-Cornell and Dillon, 2001) discuss
statistical analyses concerning general Space Shuttle risk management for the pre-Columbia
era, and detail both lapses and advancements in the use of probabilistic risk analyses (PRA).
We note here that probabilistic risk analysis covers the quantification of failure probabilities
in the absence of large volumes of data for supporting analyses, but is not directly related to
the post-design real-time monitoring phase for the detection of anomalous events. The
primary author of the previously cited study (Pat-Cornell & Dillon, 2001) has long been
considered a respected expert in the PRA field, and is often sought by NASA for opinions,
judgment, and analyses on such issues. In a study by this same author (Pat-Cornell and
Fischbeck, 1994) on the risk analyses of tiles on the space shuttle which predated the Columbia
tragedy by almost a decade, it was found that the use of PRA admitted a conclusion that
almost foretold the events of the eventual demise of the Space Shuttle Columbia.
Noted for the advocacy on the use of Bayesian statistics as opposed to classical frequentist
statistics within PRA, it is often shown in Pat-Cornells studies that a quantitative modelbased formulation requires data gathered from a variety of sources. In the case of frequentist
statistics, this requires a sufficient volume of data in order to generate a reasonably accurate
estimate of risk and confidence. The use of Bayesian analysis is more apropos to the
situation of developing risk estimates for novel engineered systems such as the Space
Shuttle, and even more so for Ares I-X where there is an inherent scarcity of data. As such,
there is the need to develop a well-informed prior distribution.
Furthermore, Pat-Cornell strongly cautions engineers against the use of a mixed approach
involving both Bayesian and frequentist statistics in order to provide for consistency, and

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

143

possibly due to previous lessons learned from experience with the application of conservative
risk estimates in tandem with probabilities. Although a pure Bayesian approach is not used in
our study, nor we do not concern ourselves with the computation and quantification of risks
and potential failures for PRA, we are motivated by another of Pat-Cornells cautions that all
too often qualitative measures are used for decision making rather than quantitative ones.
We do not present these ideas as a preamble for the investigation of the development of
such PRAs in this chapter, but rather for the monitoring of systems ostensibly that PRA
served to inform at the design and post-design certification stages. However, areas where
these previous studies may relate to ours are in the establishment of reasonably accurate
failure probabilities to inform evaluation of monitoring algorithms, and the use of a modelbased approach in tandem with a data-driven approach for monitoring. A good overview of
the need to establish a formal relationship between design and monitoring for space
applications has been discussed thoroughly in previous work (Tumer, 2005).
We note that PRA relates to design and risk assessment, whereas the use of data driven
statistical analyses performed in this study assumes that appropriate measures have already
been taken in advance in order to compute and quantify probabilities of failure. As such,
complementary maximum allowable probabilities of false alarm and missed detection are
assumed to have already been established and available. These two metrics are the essence
of the analysis we present in this chapter.
We also note as described previously that a model-based approach (TEAMS) is used in the
same architecture as a data-driven approach. This is still not applicable nor indicative of an
attempt to develop a PRA. The framework presented here relates to the monitoring of the
designed systems rather than the assessment of risk. There is evidence that the use of a
Bayesian approach has been embedded within this model-based tool (Tu et al., 2006),
although these aspects of the system are not currently being used for monitoring in this
particular demonstration or deployment. However, the availability of such functionality is a
general step in the right direction nonetheless. Pat-Cornell alludes to the use of a Bayesian
approach as being of paramount importance for inclusion in PRA, but it arguably has
similar merit for monitoring. Even though the full Bayesian updating aspects of TEAMS are
not currently being deployed for monitoring purposes, the option to include MTTF (MeanTime-To-Failure) estimates based upon the results of PRA exists in TEAMS-RT, the real-time
counterpart to our chosen model-based approach. These estimates can be used to calculate
the probability of each suspect failure mode.
Our primary concern in this chapter, however, is with an optimization, evaluation, and
performance assessment of candidate data-driven monitoring algorithms for the relevant
platform. As such, we note that there have been other related studies e.g. (Hart, 1990) on the
statistical analyses of data related to the exact subsystem that we are considering here.
However, we have found very little indication that the results of these studies were adopted
and implemented in any fashion. The mathematical sophistication of these studies was most
likely well beyond the capability of practical implementation at the time, as they were
performed nearly two decades ago. However, advancements in the use of ROC analysis that
have recently gained traction or application in the aerospace IVHM community have
motivated its current use for space-based propulsion systems of the exact type analyzed in
previous work of the current author (Martin, 2007); (Martin et al., 2007). Furthermore, ROC
analysis is a much more straightforward and broadly applicable tool used for statistical
analysis than in the previously cited work (Hart, 1990), and more useful for the design of
alarm and monitoring systems.

www.intechopen.com

144

Aerospace Technologies Advancements

3. Approach
Ostensibly, operating procedures for the thrust vector control (TVC) subsystem on Ares I-X
used to support gimbaling of the solid rocket booster are also patterned after operations of
the Space Shuttle. As such, for the purposes of our analysis here we consider two primary
phases of operation of the TVC subsystem based upon Space Shuttle operations. Testing of
the TVC in the VAB (Vehicle Assembly Building) often occurs anywhere from several weeks
to a few months prior to the point at which the vehicle is fully assembled and ready to be
rolled out on the MLP (Mobile Launch Platform) to the pad for launch. Launch countdown
begins while the assembled vehicle is already rolled out to the launch platform; however,
occasionally certain tests that normally occur in the VAB are also performed at the pad.
To simplify the distinction between the two main phases of operation, we will only consider
tests of the TVC subsystem when the vehicle is in the VAB, and when the fully assembled
vehicle is at the pad after launch countdown has commenced. Data from previous Space
Shuttle launches has been gathered from both phases, and as such both will be considered
viable candidates for analysis using the methods described previously. The availability of
the data during each phase and its ability to meet certain requirements will determine the
choice of a particular phase of operation as the primary test phase for evaluation of datadriven algorithms. Specifically, the requirement to be met is the availability of a sufficient
amount of data to undergo a formal cross-validation procedure that can be processed in a
reasonable amount of time. A priori knowledge on computational complexity of each datadriven algorithm will aid in determining if they can be processed in a reasonable amount of
time, given a fixed volume of data that sufficiently characterizes nominal behavior spanning
multiple launches.
Finally, all data must be pre-processed and transformed into a standard format fit for
utilization of all data-driven algorithms. This pre-processing can be decomposed into two
basic steps. The first step requires resampling of data that is recorded only when parameter
values change or that are derived from different data sources. This resampling technique is
performed by holding the last known value, resulting in a new data file where each row
corresponds to a single time step, and each column corresponds to a distinct parameter,
regardless of its origin. Thus, the data is merged into one aggregate matrix representation.
The second step necessitates normalization of all continuous-valued parameters so that any
potential algorithmic biases resulting from parameters that have drastically different ranges
are effectively mitigated. If any discrete-valued parameters or continuous parameters that
have fixed values exist in the dataset, they will bypass this normalization step in order to
prevent potential singularities in algorithmic processing. For the purposes of this chapter,
unless otherwise stated it will be assumed henceforth that all data has been processed and
transformed into this standard format.
3.1 IMS (Inductive Monitoring System)
The Inductive Monitoring System (IMS) is a distance-based anomaly detection tool that uses
a data driven technique called clustering to extract models of normal system operation from
archived data. IMS works with vectors of data values as described in the previous section.
During the learning process, IMS analyzes data collected during periods of normal system
operation to build a system model. It characterizes how the parameters relate to one another
during normal operation by finding areas in the vector space where nominal data tends to
fall. These areas are called nominal operating regions and correspond to clusters of nearby,

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

145

similar points found by the IMS clustering algorithm. IMS represents these nominal
operating regions as hyper-boxes in the vector space, providing a minimum and maximum
value limit for each parameter of a vector contained in a particular hyper-box. These hyperbox cluster specifications are stored in a knowledge base that IMS uses for real-time telemetry
monitoring or archived data analysis. Figure 1a shows an overview of the IMS method.

Fig. 1a. Inductive Monitoring System (IMS) Overview

3.1.1 IMS learning process
In general, the number and extent of nominal operating regions created during the IMS
learning process is determined by three learning parameters: the maximum cluster radius
can be used to adjust the size and number of clusters derived from a fixed number of
training data points, the initial cluster size is used to adjust the tolerance of newly created
nominal operating regions, and the cluster growth percent is used to adjust the percent
increase in size of a nominal operating region when incorporating new training data vectors.
More specifically, the learning algorithm builds a knowledge base of clusters from
successively processed vectors of training data. As such, the clustering approach is
incremental in nature, which distinguishes it from well-known methods such as k-means
clustering where the resulting clusters are independent of the ordering of the vectors. With
the processing of each new training data vector, the distance from this new vector to the
centroid of the nearest cluster in the knowledge base is computed. If this distance is below a
pre-specified value, the maximum cluster radius, the new vector is summarily
incorporated into that cluster. The upper or lower limits for each affected dimension of the
cluster are expanded respectively according to the cluster growth percent parameter to
reflect the inclusion of the new vector. This incremental, inductive process gives IMS an
advantage over other clustering methods such as k-means, since it tends to group
temporally related points during the learning process. The grouping of temporally related

www.intechopen.com

146

Aerospace Technologies Advancements

points may also aid in discovering distinct system operations, which makes IMS more
amenable to the specific goal of monitoring time series data for system operations.
The cluster growth percent parameter is used to adjust the learning rate. It establishes a
fixed growth percentage difference for expansion of each dimension when updating
previously formed clusters. This cluster growth percent learning parameter is therefore
clearly proportional to the learning rate, due to the increased number of training data points
that will be assigned to each new cluster per iteration for higher values of the cluster
growth percent parameter. Naturally, the number of clusters in the knowledge base for a
given training data set will increase as the maximum cluster radius and cluster growth
percent values are decreased. Therefore, an inverse relationship between the maximum
cluster radius and the number of clusters in the knowledge base exists. This dependence can
be exploited to regulate the final size of the knowledge base in order to accommodate
resource limitations in the computers running IMS, and to optimize selected metrics.
If the distance between a newly processed vector and the centroid of the nearest cluster in
the knowledge base is above the pre-specified maximum cluster radius value, a new
cluster is created. The formation of a new cluster is accomplished by creating a hyper-box
whose dimensions are based upon forming a window around each element of the new
training data vector. The window is defined by introducing the initial cluster size
parameter which is used to adjust the learning tolerance. This initial cluster size learning
parameter represents a fixed percentage of the value for each dimension of the new training
vector and as such relates directly to the size of newly established clusters. The initial
cluster size and cluster growth percent learning parameters also act as buffers which
enable a provisional allowance for manufacturing sensor tolerances and for sensors that
may have suffered from deterioration due to wear. Furthermore, these learning parameters
provide increased coverage to compensate for training data that may not fully characterize
the nominal performance envelope.
3.1.2 IMS monitoring process
During the monitoring operation, IMS reads and normalizes real-time or archived data
values, formats them into the predefined vector structure, and searches the knowledge base
of nominal operating regions to see how well the new data vector fits the nominal system
characterization. After each search, IMS returns the distance from the new vector to the
nearest nominal operating region, called the composite distance. Data that matches the
normal training data well will have a composite distance of zero. If one or more of the data
parameters is slightly outside of expected values, a small non-zero result is returned. As
incoming data deviates further from the normal system data, indicating a possible
malfunction, IMS will return a higher composite distance value to alert users to the
anomaly. IMS also calculates the contribution of each individual parameter to the composite
deviation, and the index of the nearest cluster to the monitored point, which can help
identify and isolate the cause of the anomaly.
3.1.3 IMS score computations
This section describes the computation of the IMS composite score and individual
contributing scores. The IMS scores are computed on a monitored data stream, and are
functions of distances to the nearest clusters which have been computed from training data.
Table 1 lists the notation used in the score calculation equations:

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

Table 1. Notation

Fig. 1b. Distance from point r2,s to cluster c in dimension p=2

www.intechopen.com

147

148

Aerospace Technologies Advancements

spondance is closer when the data was normalized over the range of the training

3.1.4 IMS complexity

Figs. 2 and 3 illustrate empirically generated timing tests for IMS runs that demonstrate its
computational complexity for datasets of increasing size. Fig. 2 illustrates only the time that
it takes to train IMS, while Fig. 3 shows the time that it takes to normalize, train, test, and
parse the results for IMS. In both figures, results are shown for data sets that have different
numbers of parameters, which aids in gaining a sense of scalability as well as complexity. In
Fig. 2, the size of the dataset containing 21 parameters was increased from a minimal size of
338 kB to its full size of 32 MB (blue dashed line). This data set was also noted to have been
based upon resampled data, as distinct from change only data. This distinction is notable
due to the fact that the resampled counterpart to a change-only data set is always of a
much larger size, and a change only dataset is sufficiently representative for use as a
training data set, both for obvious reasons. The resampled version is only necessary when
performing monitoring and subsequent analyses.
In Fig. 2, both resampled and change only versions of a dataset containing 164 parameters
are also shown. They are varied in size from a minimal ~ 6 MB to 218 MB (just over onethird the full size of the resampled dataset) and from 2.64 MB to 250 MB (the full size of the
change only dataset), which are shown with solid blue dots and the solid blue line,
respectively. The length of time spanned by the actual dataset (real-time) is shown in green,

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

149

and appears nearly as a vertical line due to the scale used for the graph, even though the
slope is finite. As such, it is evident that for all timing tests shown in both Figs. 2 and 3 (in
blue), IMS runs nearly in linear time and far faster than real-time, even when accounting for
the time that it takes to normalize, train, test, and parse the results for IMS (In Fig. 3). It is
also evident that the number of parameters in the dataset processed by IMS scales in a
multiplicative fashion (as well as the size of the dataset, implicitly, due to aforementioned
linear time characterization). However, for the purposes of performing a thorough analysis
that involves optimization and an n-fold cross validation, even a linear time complexity
needs to be considered carefully.

Fig. 2. IMS Training Complexity

To see how quickly time can add up, take as an example a 10-fold cross validation and
optimization performed on 5-hr training and validation datasets with 21 parameters. If we
base the total processing time for each point within a 150 point optimization grid on Fig. 3, it
becomes clear that it would exceed ~ 30 sec x 10 cases X 150 optimization points (>12 hrs), in
addition to the time for monitoring auxiliary fault data. If the validation dataset used to
monitor auxiliary fault data consists of 2 nominal and 2 fault scenarios, and each scenario
spans 5 hrs, then the time for monitoring fault data for 2 nominal and 2 fault scenarios
exceeds ~ 4 sec X 10 cases X150 optimization points X 4 scenarios, based upon Fig. 2 (using
training time as a rough estimate for monitoring time). The total estimated time amounts
to almost 20 hrs. of computational effort. And even so, these rough computations do not
account for the fact that for file sizes above the max 250 MB used to create these plots, the
number of clusters found by IMS may increase substantially, also adding significantly to the
computational cost, possibly running in near real-time or even above real-time for very large
files.

www.intechopen.com

150

Aerospace Technologies Advancements

Fig. 3. IMS Normalization, Training, Testing, and Results Parsing Complexity

Evidence of this variation with number of clusters is shown in Fig. 4. Here, training
/validation data file sizes are fixed for the entire experiment. The experiment implicitly
varies the number of clusters with each subsequent training/validation run via adjustment
of IMSs maximum cluster radius parameter. The x axis represents the number of
sequential IMS training/validation set completions, the y axis on the left represents the date
of completion, and the y axis on the right represents the number of clusters. As in Fig. 3, this
involves IMS training, testing, and the overhead involved in wrapper code written for
generation of statistics for optimization, which includes parsing of the results. The reason
sequentially processing training/validation runs appears sublinear is due to the fact that the
maximum cluster radius value is being adjusted from a very small value to 1. Recall that
there is always an inverse relationship between the maximum cluster radius and the
number of clusters. As such, we implicitly decrease the number of clusters per IMS
knowledge base for each subsequent run from 976 to 18, causing the time for completion to
be shortened decrementally.
For files of an increasing size which seeded the experimental results shown in Figs. 2 and 3,
the effect seen in Fig. 4 may bias the attempt to control the experiment by holding the
maximum cluster radius parameter fixed. When the file size is increased for the experiments
yielding the results shown in Figs. 2 and 3, the number of clusters naturally increases due to
the greater chance of visiting areas in the hyperspace that contain vectors previously unseen
by the current IMS cluster population. As such, a similar sublinear effect as seen in Fig. 4
may result for an experiment in which the file size is decreased but the maximum cluster
radius parameter is held fixed. Conversely, a limiting effect is apparent when the number of

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

151

clusters increases as a side effect of increasing the file size, implying that IMS will run in
near linear time as file size increases (as is evident in Figs. 2 and 3). However, this limiting
linear time effect does not imply that IMS will run in real time for an arbitrarily large
number of clusters. In fact, as mentioned previously, it is possible for IMS to run above realtime for very large files, since we know that IMS scales in a multiplicative fashion with the
size of the dataset. As such, an increase in the number of clusters resulting from an increase
in file size will yield a substantially greater computational burden. In the previous example,
rough estimates were provided that did not account for these effects.

Fig. 4. IMS Complexity and corresponding number of clusters

3.2 Orca
Orca is a software tool that uses a nearest neighbor based approach to outlier detection
which is based upon the Euclidean distance metric. It uses a modified pruning rule that
allows for increased computational efficiency, running in near linear time as the number of
top score outlier points selected for evaluation decreases. More information on this
algorithm and some of its applications can be found in previous work (Bay and
Schwabacher, 2003); (Schwabacher, 2005). This algorithm outputs a total score which
represents the average distance to the nearest k neighbors in the multi-dimensional feature
space containing all of the variables. It also outputs the contribution of each variable to this
score in order to show which variables cause each outlier to be classified as such.
Unlike IMS, Orca requires no training due the inherent comparative nature of the algorithm.
Ostensibly, this should reduce the total computational burden. However, the greatest

www.intechopen.com

152

Aerospace Technologies Advancements

computational advantages of using Orca are reaped only when applying a modified
pruning rule for a reduced number of outlier points based upon having the highest score
magnitude selected for evaluation. An implicit requirement to use all available points exists,
to aid in the construction of an unbiased ROC curve representative of all score magnitudes
and with full resolution. Therefore, Orca loses is computational advantage and defaults to a
standard k-nn (nearest neighbor) algorithm that scales quadratically. A two-dimensional
example is provided in Fig. 5 in order to illustrate the basic idea behind the k-nn algorithm.
The red circle represents the point being monitored, and the average Euclidean distance to a
fixed number of nearest neighbors, k, is used to compute a composite anomaly score.
Isolation of the outlier point becomes possible by using this technique as the monitored
point is indexed across the entire population, as is evident in the diagram.

Fig. 5. Nearest Neighbor Approach to Anomaly Detection

Fig. 6. Orca Normalization, Testing, and Results Parsing Complexity

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

153

Empirical evidence of quadratic complexity is shown in Fig. 6, which also illustrates the time
required for normalization, application of Orca, and results parsing. This graph merits direct
comparison to Fig. 3 showing complexity results for IMS, due to use of the same datasets and
ranges in size of the datasets. Clearly, the computational complexity of Orca is far greater than
that of using IMS. Furthermore, for datasets spanning a large period of time, it is readily
apparent that the algorithm will eventually no longer be able to run faster than real time.
3.3 One-Class Support Vector Machine (SVM)
The one-class support vector machine is a very specific instance of a support vector machine
which is geared for anomaly detection. The generic support vector machine (SVM) can be
used to classify data in multiple dimensions by finding an appropriate decision boundary.
Like neural networks, support vector machines perform classification using nonlinear
boundaries, however they go one step beyond neural networks by finding the boundaries
that provide the maximum margin between different classes of data. Additionally, using the
support vector machine one can map data from a lower dimensional space that is not
linearly separable to a higher (even infinite-dimensional) space where the data are linearly
separable by a hyperplane. This is performed by using what is commonly known in
machine learning as the kernel trick, when using SVMs. A kernel function is chosen to
map the data from the lower-dimensional space to the higher-dimensional space. It can be
chosen arbitrarily so as to best suit the data and at the same time reduce the computational
burden involved with generating the mapped values by direct evaluation. However, for our
purposes, we choose the Gaussian radial basis function or kernel, given by Eqn. 9, which is
the most widely used in the application of one-class SVMs.
(9)
K is the kernel function, and the training data vectors are given by xi and xj. The kernel
width is given by , and is used to control the distribution of the kernel function around the
training data point. The magnitude of this value also often has a direct effect on the
algorithms speed and complexity. Support vectors correspond to those data points that
lie along the margin or closest to it. The maximum margin between classes is found by
solving a quadratic optimization problem.
The one-class SVM differs from the generic version of the SVM in that the resulting
quadratic optimization problem includes an allowance for a certain small predefined
percentage of outliers, , making it suitable for anomaly detection. As shown in Fig. 7, these
outliers lie between the origin and the optimal separating hyperplane. All the remaining
data fall on the opposite side of the optimal separating hyperplane, belonging to a single,
nominal class, hence the terminology one-class SVM. The SVM outputs a score that
represents the distance from the data point being tested to the optimal hyperplane. Positive
values for the one-class SVM output represent normal behavior (with higher values
representing greater normality) and negative values represent abnormal behavior (with
lower values representing greater abnormality). For subsequent analyses and in the results
section, the negative of this score will be used in order to establish a frame of reference
commensurate with ROC analysis and for comparison to other algorithms. More technical
details on the one-class SVM are available in previously published studies (Das et al., 2007);
(Cohen et al., 2004).

www.intechopen.com

154

Aerospace Technologies Advancements

Fig. 7. Illustration of One-Class Support Vector Machine Framework

The one-class SVM differs from the other methods discussed in this paper because it
determines whether or not a point is an outlier based on the distance of the point to a
separating hyperplane in a feature space induced by a kernel operator, whereas most of the
other methods rely on an analysis of the data in the original data space. For the one-class
SVM, a single hyperplane separates the nominal data from the origin. Thus, for a system
which undergoes nominal mode changes during its operation, all such changes will be
characterized as nominal with a single hyperplane. Orca and IMS, on the other hand,
characterize the anomalousness of a point based on local characteristics within the data
space. This quality can make those algorithms more robust to significant mode changes
compared with the one-class SVM.
The small predefined percentage of outliers, , is typically selected a priori and used to
inform optimal selection of a value for the kernel width, , as is shown in Fig. 8. This is
performed by finding the correct training classification rate as a function of the kernel
width, and finding the value of that corresponds to 1-. It is often easy to find this value
within a reasonable level of tolerance. In these cases, the kernel width, , corresponding to
the first exceedance of 1- is used as the optimal value. However, when using a very small
value of , it may be infeasible to find a training misclassification rate that will achieve this
value (1-). This puts a lower bound on the predefined percentage of outliers that can be
selected. As such, various candidates can be tested for feasibility in order to determine the
lower bound on . For each of these candidates, a corresponding relevant metric can be
computed, in addition to the corresponding optimal kernel width. In this case, the relevant
test metric of interest is the AUC value. A value of that has been determined to be feasible,

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

155

and corresponds to the highest AUC value will thus be used for selection of the optimal
kernel width.

Fig. 8. Illustration of Selection of Optimal Kernel Width

The complexity of the one-class SVM algorithm is also of interest. As such, we provide an
empirical complexity analysis shown in Fig. 9 to complement Figs. 2-4 and 6, which
illustrate the complexity for IMS and Orca. Similar to Orca in Fig. 6, it is evident that the
one-class support vector machine algorithm yields a superlinear complexity shown in Fig. 9.
Furthermore, for the smaller dataset containing 21 resampled parameters, we compare the
time for training only (shown as a dotted line), to the time required for normalization,
training and validation (shown as a dash-dot line), which adds negligible time. It is also
interesting to note that there is no noticeable multiplicative increase in complexity with the
processing of additional parameters as opposed to Orca and IMS (cf. 21 resampled
parameters to 164 change-only parameters), which is ostensibly due to the use of the kernel
trick. However, regardless of this slight advantage, again the computational complexity of
the one-class SVM algorithm is far greater than that of IMS, and on par with that of Orca. As
seen in Fig. 9 , for datasets spanning a period of time greater than 7 hrs., it is readily
apparent that the OCSVM algorithm is no longer able to run faster than real time.

4. Validation and evaluation criteria

In order to motivate a detailed discussion of the evaluation criteria, we will first need to
introduce the means by which we will validate the algorithms discussed thus far.
Specifically, we need to outline the type and volume of data required for validation. Recall

www.intechopen.com

156

Aerospace Technologies Advancements

Fig. 9. One-Class Support Vector Machine Computational Complexity

the requirement to process a sufficient amount of data to undergo a formal cross-validation
procedure that can be processed in a reasonable amount of time. The computational
complexity of each data-driven algorithm is now well established, which facilitates the
selection of various candidate cross-validation experiments. Again, for the purposes of our
analysis here we will only consider two primary phases of operation of the TVC subsystem
based upon meeting these specified data requirements. The two phases of interest involve
both testing of the TVC in the VAB, and when the fully assembled vehicle is at the pad after
launch countdown has commenced. For each of these phases, there are specific periods of
time in which respective tests of interest or important elements of the operational sequence
are conducted. Certain pre-specified failure modes from a FMECA (Failure Modes, Effects
and Criticality Analysis) have been identified as potential adverse events that could occur
during these critical periods of time.
Table 2 provides a summary of each failure mode under consideration for each phase of
operation. The periods of time spanned by tests corresponding to each of these failure
modes are also provided, which aid important considerations of computational complexity.
For the last three row entries in Table 2, the period of time spanned by the tests are largely
dependent upon the specific flight, and are also driven by acquisition of a sufficient amount
of data to accurately characterize nominal behavior in order to prevent false alarms. In light
of this fact, the use of a 25 Hz sampling rate, and the example of how quickly time
complexity for even the least burdensome algorithm adds up, we will use the first two row
entries in Table 2 as the basis of experiments and results to be subsequently presented.
Another advantage of using the two failure scenarios at the pad is due to the fact that they
occur during the same time period within T 1 min prior to launch, thus simplifying the
experimental construction and allowing for fewer scenarios to investigate independently.

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

Operational
Phase
Pad
Pad
VAB

VAB

VAB

VAB

157

Period of
Time Spanned
FSM (Fuel Supply Module) Pressure Drop due Within T 1 min
prior to launch
to N2H4 (Hydrazine) Leak
Hydraulic Fluid Reservoir Level Drop due to
Within T 1 min
hydraulic fluid leak
prior to launch
Actuator Stuck during Actuator positioning
2.5 min test in VAB
test
Within 34 minute
FSM (Fuel Supply Module) Pressure Drop due period after
to N2H4 (Hydrazine) Leak
calibration test in
VAB
Within 10 min
Hydraulic Fluid Reservoir Level Drop due to
period during TVC
hydraulic fluid leak
actuator tests in VAB
Within 25 min
Hydraulic pumping unit overtemperature
period during tests in
failure
VAB
Failure Mode

Table 2. Failure Mode Summary

Due to the absence of real data to use for validation and testing for the two candidate failure
modes under consideration in Table 2, they must be simulated using existing means.
Increasing the range of fidelity of available failure simulations is an ongoing process. In
previous work, (Schwabacher et al., 2009) a high fidelity physics-based simulation of a leak
in a liquid propulsion-based J-2X engine was used to validate data-driven algorithms.
However, such a simulator does not currently exist for relevant components of the TVC
subsystem. In future work, we plan to develop high fidelity physics-based simulations of a
leak in the fuel supply module (FSM) of the TVC subsystem, which is a spherical tank
containing liquid hydrazine (N2H4) pressurized by GN2 (gaseous nitrogen). However, for
our purposes here we have chosen to use a rather low fidelity simulation for the two
scenarios. The simple use of linearly decreasing ramps from nominal operating conditions to
off-nominal values for parameters specific to each failure mode will be used, given
predefined rates of degradation and preselected off-nominal values.
The time of failure injection within the T- 1 min period prior to launch is selected to leave a
significant nominal fraction of the period at from T -1 min until injection of the failure,
which occurs at T-28.12 sec for the hydrazine leak at the pad, levelling off at its off-nominal
value at T-15 sec. For the hydraulic fluid reservoir leak at the pad, the failure is injected at T13 sec and levels off at its nominal value at T-12 sec. However, it is often useful to
supplement our validation data with additional cases of purely nominal behavior as well. In
our case, we will use a concatenation of nominal data from T-1 min to launch, and two
additional failure datasets representing the scenarios described above. Now that we have
established the basis for constructing the validation data sets, it is of interest to describe how
the algorithms will be optimized and evaluated with the use of ROC curve analysis.
The ROC curve essentially plots the true positive rate against the false alarm rate for all
possible threshold values, as shown in Fig. 10. It therefore can be used as a design tool in
order to select an alert threshold according to pre-established requirements for minimum

www.intechopen.com

158

Aerospace Technologies Advancements

missed detection and/or false alarm rates. However, in order to compute true positive and
false alarm rates, it is necessary to obtain a ground truth representation for each
monitored example. In this case, an example can represent an individual validation flight,
or a single time point. Here we will compute false alarm and missed detection rates based
upon the classification of each individual time point, rather than on a flight-by-flight basis,
as was performed in previous studies (Martin et al., 2007).
In order to compute true positive and false alarm rates with a reasonable level of accuracy,
there is a need to obtain a statistically significant number of labeled examples, both nominal
and anomalous. The availability of nominally classified examples often far exceeds the
availability of anomalously classified examples. The deficit of the latter drove our efforts to
simulate faults, and hence to use classified time points as monitored examples in lieu of
individual validation flights for construction of the ROC curve. For our purposes here, the
classified time points represent the ground truth, and will be based upon the times of
failure injection specified previously, for each failure scenario.
One advantage of using ROC curve analysis is in its inherent robustness against the use of
skewed distributions between the populations of nominally and anomalously categorized
examples. Thus, as long as the population of failure examples is statistically significant, it
does not need to be on par with the population of nominal examples. Furthermore, unlike
other alert threshold selection techniques, the ROC curve may also be used for optimized
selection of algorithmic parameters. This can be performed by using the area under the ROC
curve, which represents overall classification discriminability. Therefore, the AUC can be
maximized with respect to free design parameters used for algorithmic tuning to optimize
and control performance during the validation stage. As such, we can ensure that the
resulting model used to perform threshold or alert selection has the best anomaly detection
capability possible. Furthermore, the AUC analysis can be used to compare the performance
of all candidate algorithms.

Fig. 10. Sample ROC Curve

As seen in Fig. 10, the area under the ROC curve has a classic increase in relation to the SNR
(Signal-to-Noise ratio). This relationship is well established, and is derived from the origins

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

159

of the ROC curve for use in radar applications dating back to WWII. Unlike the SNR, it is
often the case that algorithmic design parameters do not have a similar
straightforward relationship to the AUC. However, this relationship is implicit in Fig. 10
due to the fact that optimization of algorithmic design parameters is performed by using the
area under the ROC curve. In a sense, the ROC curve is "tuned" by attempting to maximize
the AUC (area under the ROC curve) and choosing the appropriate algorithmic design
parameter(s) to allow for maximum predictive capability. This can be thought of as choosing
the ROC curve with the highest SNR based upon such parameters. In this case theoretically
the "signal" might loosely be thought of as the variability of anomalous behavior, and the
"noise" might be thought of as the variability of nominal operations.

5. Results and discussion

As previously discussed, a formal cross-validation procedure will be used to perform
optimization and alert threshold selection. Doing so will help to prevent overfitting, as is
often the case when using a single training and validation dataset. A k-fold cross-validation
will be used, in which k-1 training examples are used, and the remaining example is used for
validation. Each of the k examples is rotated and used one time for use as a validation set, so
that there are k sets of statistics rather than a single one. The candidate training/validation
dataset partitions will span 7 flights (STS-117, STS-107, STS-112, STS-113, STS-114, STS-120,
and STS-122) containing 30 continuous-valued parameters, and a 7-fold cross-validation
procedure will be used. Resulting AUC and ROC curves will be averaged over the results of
the 7-fold cross-validation. The optimization parameters used and the respective optimized
values derived from the average AUC across all 7 folds for each of the candidate data-driven
algorithms are summarized in Table 3.
Algorithm

IMS

Orca
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM
OCSVM

Optimization
Parameter

Maximum cluster
radius
(MAX_INTERP)
k (number of
nearest neighbors)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)
(kernel width)

Table 3. Optimization Parameters

www.intechopen.com

Max. of
averaged AUC
Value Achieved

Hyperparameter
Values

.01894
(403 clusters)

.9807

Initial cluster size

(INIT_TOLERANCE) =
0.01
Cluster growth percent
(EXTRAP_PERCENT) =
0.05

0.91077

N/A

10.8081
2.121
10.8081
2.121
0.74531
2.121
5.9636
5.9636
3.5565
infeasible

0.92956
0.93072
0.90416
0.9051
0.94666
0.90664
0.90529
0.90525
0.89681
infeasible

Optimized
Value

= 0.2289
= 0.2
= 0.1
= 0.08
= 0.075
= 0.07
= 0.05
= 0.01
= 0.001
= 0.0001

160

Aerospace Technologies Advancements

Additionally, the optimized maximum AUC values achieved by averaging across all 7 folds,
and values of any supplementary hyperparameters that were used for each algorithm are
provided in Table 3. Fig. 11 illustrates the average AUC values across all 7 folds for each of
the candidate data-driven algorithms as a function of each of their design parameters. In
addition, quasi-confidence bounds representing the maximum and minimum AUC values
across all 7 folds are provided. In the lower right panel of Fig. 11, the maximum of the
averaged AUC value achieved that lies in a feasible range is demarcated, even though it is
clearly not the maximum over the entire range of kernel width values shown. The range of
feasible kernel width values for the row in Table 3 highlighted in red represents any value
greater than the optimized value shown, and corresponds to the lower right panel in Fig. 11.
Both the highlighted row and the lower right panel of Table 3 and Fig. 11 respectively
represent the maximum achievable AUC averaged across all 7 folds for the OCSVM
algorithm. Thus, the optimized kernel width lies on the very border of the feasible range.
Note that the last row of Table 3 indicates that a lower bound on the predefined percentage
of outliers, min, has been identified for any kernel width.

Fig. 11. AUC optimization with confidence intervals

Conflating the values presented in Table 3 with the graphs shown in Fig. 11, it becomes clear
that IMS outperforms the other two algorithms, even when the hyperparameters for IMS are
held fixed and not implicitly used to inform the optimization problem, as was performed
with the one-class SVM. However, what may not be as clear from Fig. 11 is that some
overlap between the confidence intervals among the algorithms exists. Specifically, for IMS,

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

161

the averaged optimized AUC value achieved falls within the range (0.9728, 0.9921), for Orca
the range for optimized AUC values is (0.8924, 0.9201), and for OCSVM the range for
optimized AUC values is (0.919, 0.9891). Therefore, there is some overlap between the
optimum ranges for IMS and OCSVM, albeit a very small overlap of 0.0163 in which SVM
might outperform IMS. Orca has no range overlap with IMS, but has an extremely small
overlap with SVM of 0.0011.
Our next and final step is to perform threshold or alert selection based upon the ROC curves
corresponding to the optimized parameters found from the previous step. Recall that this
allows for alert threshold selection based upon specified minimum false alarm and/or
missed detection rates with a robust anomaly detection capability and the best classification
discriminability. Fig. 12 shows examples of the optimized ROC curves averaged over all 7
folds and their minimum/maximum confidence interval counterparts. Additionally,
corresponding alert thresholds have been selected according to an established maximum
allowable false positive rate of 0.01, and the results are shown in the legend.

Fig. 12. ROC Curve Comparison and Alert Threshold Selection

The slight variation in ranges shown in the legend differ from the optimized AUC ranges
previously provided due to the revised manner in which they have been computed. The
minimum and maximum ROC curves were constructed in a modified manner in order to
prevent any excursions of the mean ROC curve outside of the min/max envelope for
visualization purposes. As such, the mean false positive rate is used consistently to construct

www.intechopen.com

162

Aerospace Technologies Advancements

ROC curves representing the min/max bounds, while the min/max true positive rate is
used respectively, resulting in different min/max AUC values than previously reported,
however these variations are negligible. Note that for a fixed false alarm rate ceiling (Pfa) of
0.01, the resulting OCSVM threshold yields a true positive rate (Pd) on par with that of IMS
(mean 0.79), and the resulting Orca threshold yields a much lower a true positive rate (Pd)
having a mean of 0.11. In due fairness to Orca, it is possible that the AUC value may have
increased beyond the range of the 200 nearest neighbors shown on the upper right panel of
Fig. 11. However, due to scalability and complexity issues, the experiment was terminated
as shown. Indeed, the complexity of running Orca with an increased number of nearest
neighbors scales quadratically when not using any pruning rules.
Finally, Fig. 13 illustrates realizations of validation flight STS-117 to which optimized
parameters and thresholds are applied.

Fig. 13. Realizations of Optimized Parameters and Thresholds Applied to Validation Flight
Fig. 13 illustrates both the nominal (red lines) and fault injected scores (blue lines) from T 1
min to launch at the pad for all algorithms. In all cases, both failure scenarios involving an
FSM pressure drop (left panels) and a hydraulic fluid reservoir leak (right panels) are
shown, using the optimized thresholds (fixed magenta colored lines) found from the ROC
analysis. The actual threshold values may vary slightly from the values specified in Fig. 12

www.intechopen.com

Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch Shuttle Operations

163

due to realizations being based on independent experiments, and as such should still fall
within the confidence bands formed in both Figs. 11 and 12. The light blue highlights
superimposed over the failure injected scores in dark blue represent the ground truth time
of failure injection and duration, so as to give a feel for the false alarm and correct detection
rates. Evidently, there is a clear bifurcation between nominal and anomalous scores for both
IMS and SVM, and for Orca the same is true although it is less apparent. As can be discerned
from Figs. 11 13, we have identified the fact that both in complexity and accuracy, IMS seems
to be the best choice among all of the algorithms investigated. However, there is some overlap
in the confidence intervals for IMS and SVM AUC values, and the alert thresholds applied for
both corresponding ROC curves yield almost identical true positive rates.

6. Conclusion and next steps

We have provided a thorough end-to-end description of the process for evaluation of three
different data-driven algorithms for anomaly detection. Through optimization of
algorithmic parameters using the AUC, we were able to choose parameters yielding the best
detection capability. The respective ROC curves corresponding to these parameters were
then used to inform alert threshold selection by enforcement of a maximum allowable false
alarm rate. It was found that IMS was the best performing algorithm when considering both
computational complexity and accuracy. However, when evaluating the results based upon
accuracy alone, the OCVSM approach is competitive with IMS due to overlapping
confidence intervals present in the accuracy results.
In subsequent research studies, we will provide results of unseen hold out test cases to
which optimized parameters and thresholds will be applied, in order to provide additional
evidence demonstrating the superiority of a particular algorithmic technique. Furthermore,
we will employ a variant of the AUC that only considers performance evaluation for
algorithmic comparison restricted to low false positive rates. A slightly modified definition
of false alarms and missed detections that accounts for pre-defined latencies and prediction
horizons will also be investigated.

7. Acknowledgements
The author would like to acknowledge the support of the Ares I-X Ground Diagnostics
Prototype activity, which was funded by the NASA Constellation Program, by the NASA
Exploration Technology Development Program, and by the NASA Kennedy Space Center
Ground Operations Project. Furthermore, the author graciously acknowledges the reviews
of Dr. Mark Schwabacher, Bryan Matthews, and Dr. Ann Patterson-Hine. The author also
extends appreciation to John Wallerius for his contribution of the subsection on IMS score
computations, and his ideas pertaining to next steps on consideration of performance
evaluation for algorithmic comparison restricted to low false positive rates. Finally, the
author acknowledges the permission to use of Figs. 7 and 8 from Dr. Santanu Das, and Fig.
1a with supporting text from David Iverson.

8. References
Bay, S.D. & Schwabacher, M. (2003). Mining distance-based outliers in near linear time with
randomization and a simple pruning rule, Proceedings of The Ninth ACM SIGKDD

www.intechopen.com

164

Aerospace Technologies Advancements

International Conference on Knowledge Discovery and Data Mining, pp. 2938, New
York, NY, 2003.
Cavanaugh, K. (2001). An integrated diagnostics virtual test bench for life cycle support,
Proceedings of the IEEE Aerospace Conference, pp. 7323573246, ISBN: 0-7803-6599-2,
Big Sky, Montana, March 2001.
Cohen, G.; Hilario, M. & Pellegrini, C. (2004). One-class support vector machines with a
conformal kernel: a case study in handling class imbalance, In: Structural, Syntactic,
and Statistical Pattern Recognition, A. Fred et al. (Eds.), pp. 850858, Springer-Verlag,
Berlin, Heidelberg, Germany.
Das, S.; Srivastava, A. & Chattopadhyah, A. (2007). Classification of Damage Signatures in
Composite Plates using One-Class SVMs, Proceedings of the IEEE Aerospace
Conference, Big Sky, MO, March 2007.
Fragola, J.R. (1996). Space shuttle program risk management, Proceedings of the International
Symposium on Product Quality and Integrity: Reliability and Maintainability Symposium,
ISBN: 0-7803-3112-5, pp. 133142, Jan 1996.
Hart, G.F. (1990). Launch commit criteria performance trending analysis, Annual Proceedings
of the Reliability and Maintainability Symposium, pp. 3641, Jan 1990.
Iverson, D.L.; Martin, R.; Schwabacher, M.; Spirkovska, L.; Taylor, W.; Mackey, R. & Castle.
J.P. (2009). General purpose data-driven system monitoring for space operations,
Proceedings of the AIAA Infotech@Aerospace Conference, Seattle, Washington, April
2009.
Mackey, R.; James, M.; Park, H. & Zak. M. (2001). BEAM: Technology for autonomous selfanalysis, Proceedings of the IEEE Aerospace Conference, Big Sky, MT, 2001.
Martin, R. (2007). Unsupervised anomaly detection and diagnosis for liquid rocket engine
propulsion, Proceedings of the IEEE Aerospace Conference, Big Sky, MT, March 2007.
Martin, R.; Schwabacher, M.; Oza, N. & Srivastava, A. (2007). Comparison of unsupervised
anomaly detection methods for systems health management using Space Shuttle
main engine data, Proceedings of the 54th Joint Army-Navy-NASA-Air Force Propulsion
Meeting, Denver, CO, May 2007.
Pat-Cornell, E. & Dillon, R. (2001). Probabilistic risk analysis for the NASA space shuttle: a
brief history and current work. Reliability Engineering & System Safety, Vol. 74, No. 3,
(Dec. 2001) pp. 345 352.
Pat-Cornell, E. & Fischbeck, P.S. (1994). Risk Management for the Tiles of the Space Shuttle.
Interfaces, Vol. 24, No. 1, (Jan. Feb. 1994) pp. 6486.
Schwabacher, M. (2005). Machine learning for rocket propulsion health monitoring,
Proceedings of the SAE World Aerospace Congress, pp. 11921197, Dallas, Texas, 2005.
Schwabacher, M. & Waterman, R. (2008). Pre-launch diagnostics for launch vehicles,
Proceedings of the IEEE Aerospace Conference, Big Sky, MT, March 2008.
Schwabacher, M; Aguilar, R & Figueroa, F. Using Decision Trees to Detect and Isolate
Simulated Leaks in the J-2X Rocket Engine, Proceedings of the IEEE Aerospace
Conference, Big Sky, MT, 2009.
Tu, H.; Allanach J.; Singh S.; Pattipati K.R. & Willett, P. (2006). Information integration via
hierarchical and hybrid Bayesian networks. IEEE Transactions on Systems, Man and
Cybernetics, Part A, Vol. 36, No. 1 (Jan. 2006) pp.1933.
Tumer, I. (2005). Design methods and practices for fault prevention and management in
spacecraft, Technical report, NASA Ames Research Center, 2005.

www.intechopen.com

Aerospace Technologies Advancements

Edited by Thawar T. Arif

ISBN 978-953-7619-96-1
Hard cover, 492 pages
Publisher InTech

Published online 01, January, 2010

Published in print edition January, 2010

Space technology has become increasingly important after the great development and rapid progress in
information and communication technology as well as the technology of space exploration. This book deals
with the latest and most prominent research in space technology. The first part of the book (first six chapters)
deals with the algorithms and software used in information processing, communications and control of
spacecrafts. The second part (chapters 7 to 10) deals with the latest research on the space structures. The
third part (chapters 11 to 14) deals with some of the latest applications in space. The fourth part (chapters 15
and 16) deals with small satellite technologies. The fifth part (chapters 17 to 20) deals with some of the latest
applications in the field of aircrafts. The sixth part (chapters 21 to 25) outlines some recent research efforts in
different subjects.

How to reference

In order to correctly reference this scholarly work, feel free to copy and paste the following:
Rodney A. Martin, Ph.D. (2010). Evaluation of Anomaly Detection Capability for Ground-Based Pre-Launch
Shuttle Operations, Aerospace Technologies Advancements, Thawar T. Arif (Ed.), ISBN: 978-953-7619-96-1,
InTech, Available from: http://www.intechopen.com/books/aerospace-technologies-advancements/evaluationof-anomaly-detection-capability-for-ground-based-pre-launch-shuttle-operations

InTech Europe

University Campus STeP Ri

Slavka Krautzeka 83/A
51000 Rijeka, Croatia
Phone: +385 (51) 770 447
Fax: +385 (51) 686 166
www.intechopen.com

InTech China

Unit 405, Office Block, Hotel Equatorial Shanghai

No.65, Yan An Road (West), Shanghai, 200040, China
Phone: +86-21-62489820
Fax: +86-21-62489821